1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-igemm/sse-load1.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <xmmintrin.h>
13 
14 #include <xnnpack/igemm.h>
15 
16 
xnn_f32_igemm_minmax_ukernel_5x8__sse_load1(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_5x8__sse_load1(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const float**restrict a,
23     const float*restrict w,
24     float*restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const float* zero,
29     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31   assert(mr != 0);
32   assert(mr <= 5);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(float) == 0);
36   assert(ks != 0);
37   assert(ks % (5 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(float) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   float* c0 = c;
44   float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52   float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53   if XNN_UNPREDICTABLE(mr < 4) {
54     c3 = c2;
55   }
56   float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
57   if XNN_UNPREDICTABLE(mr <= 4) {
58     c4 = c3;
59   }
60 
61   do {
62     __m128 vacc0x0123 = _mm_load_ps(w);
63     __m128 vacc0x4567 = _mm_load_ps(w + 4);
64     __m128 vacc1x0123 = vacc0x0123;
65     __m128 vacc1x4567 = vacc0x4567;
66     __m128 vacc2x0123 = vacc0x0123;
67     __m128 vacc2x4567 = vacc0x4567;
68     __m128 vacc3x0123 = vacc0x0123;
69     __m128 vacc3x4567 = vacc0x4567;
70     __m128 vacc4x0123 = vacc0x0123;
71     __m128 vacc4x4567 = vacc0x4567;
72     w += 8;
73 
74     size_t p = ks;
75     do {
76       const float* restrict a0 = a[0];
77       assert(a0 != NULL);
78       if XNN_UNPREDICTABLE(a0 != zero) {
79         a0 = (const float*) ((uintptr_t) a0 + a_offset);
80       }
81       const float* restrict a1 = a[1];
82       assert(a1 != NULL);
83       if XNN_UNPREDICTABLE(a1 != zero) {
84         a1 = (const float*) ((uintptr_t) a1 + a_offset);
85       }
86       const float* restrict a2 = a[2];
87       assert(a2 != NULL);
88       if XNN_UNPREDICTABLE(a2 != zero) {
89         a2 = (const float*) ((uintptr_t) a2 + a_offset);
90       }
91       const float* restrict a3 = a[3];
92       assert(a3 != NULL);
93       if XNN_UNPREDICTABLE(a3 != zero) {
94         a3 = (const float*) ((uintptr_t) a3 + a_offset);
95       }
96       const float* restrict a4 = a[4];
97       assert(a4 != NULL);
98       if XNN_UNPREDICTABLE(a4 != zero) {
99         a4 = (const float*) ((uintptr_t) a4 + a_offset);
100       }
101       a += 5;
102 
103       size_t k = kc;
104       do {
105         const __m128 vb0123 = _mm_load_ps(w);
106         const __m128 vb4567 = _mm_load_ps(w + 4);
107         w += 8;
108 
109         const __m128 va0 = _mm_load1_ps(a0);
110         a0 += 1;
111         const __m128 va1 = _mm_load1_ps(a1);
112         a1 += 1;
113         const __m128 va2 = _mm_load1_ps(a2);
114         a2 += 1;
115         const __m128 va3 = _mm_load1_ps(a3);
116         a3 += 1;
117         const __m128 va4 = _mm_load1_ps(a4);
118         a4 += 1;
119 
120         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
121         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
122         vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
123         vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
124         vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
125         vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
126         vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123));
127         vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567));
128         vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123));
129         vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567));
130         k -= sizeof(float);
131       } while (k != 0);
132       p -= 5 * sizeof(void*);
133     } while (p != 0);
134 
135     const __m128 vmax = _mm_load_ps(params->sse.max);
136     vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
137     vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
138     vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
139     vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
140     vacc4x0123 = _mm_min_ps(vacc4x0123, vmax);
141     vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
142     vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
143     vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
144     vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
145     vacc4x4567 = _mm_min_ps(vacc4x4567, vmax);
146 
147     const __m128 vmin = _mm_load_ps(params->sse.min);
148     vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
149     vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
150     vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
151     vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
152     vacc4x0123 = _mm_max_ps(vacc4x0123, vmin);
153     vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
154     vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
155     vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
156     vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
157     vacc4x4567 = _mm_max_ps(vacc4x4567, vmin);
158 
159     if XNN_LIKELY(nc >= 8) {
160       _mm_storeu_ps(c4, vacc4x0123);
161       _mm_storeu_ps(c4 + 4, vacc4x4567);
162       c4 = (float*) ((uintptr_t) c4 + cn_stride);
163       _mm_storeu_ps(c3, vacc3x0123);
164       _mm_storeu_ps(c3 + 4, vacc3x4567);
165       c3 = (float*) ((uintptr_t) c3 + cn_stride);
166       _mm_storeu_ps(c2, vacc2x0123);
167       _mm_storeu_ps(c2 + 4, vacc2x4567);
168       c2 = (float*) ((uintptr_t) c2 + cn_stride);
169       _mm_storeu_ps(c1, vacc1x0123);
170       _mm_storeu_ps(c1 + 4, vacc1x4567);
171       c1 = (float*) ((uintptr_t) c1 + cn_stride);
172       _mm_storeu_ps(c0, vacc0x0123);
173       _mm_storeu_ps(c0 + 4, vacc0x4567);
174       c0 = (float*) ((uintptr_t) c0 + cn_stride);
175 
176       a = (const float**restrict) ((uintptr_t) a - ks);
177       nc -= 8;
178     } else {
179       if (nc & 4) {
180         _mm_storeu_ps(c4, vacc4x0123);
181         _mm_storeu_ps(c3, vacc3x0123);
182         _mm_storeu_ps(c2, vacc2x0123);
183         _mm_storeu_ps(c1, vacc1x0123);
184         _mm_storeu_ps(c0, vacc0x0123);
185 
186         vacc4x0123 = vacc4x4567;
187         vacc3x0123 = vacc3x4567;
188         vacc2x0123 = vacc2x4567;
189         vacc1x0123 = vacc1x4567;
190         vacc0x0123 = vacc0x4567;
191 
192         c4 += 4;
193         c3 += 4;
194         c2 += 4;
195         c1 += 4;
196         c0 += 4;
197       }
198       if (nc & 2) {
199         _mm_storel_pi((__m64*) c4, vacc4x0123);
200         _mm_storel_pi((__m64*) c3, vacc3x0123);
201         _mm_storel_pi((__m64*) c2, vacc2x0123);
202         _mm_storel_pi((__m64*) c1, vacc1x0123);
203         _mm_storel_pi((__m64*) c0, vacc0x0123);
204 
205         vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
206         vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
207         vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
208         vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
209         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
210 
211         c4 += 2;
212         c3 += 2;
213         c2 += 2;
214         c1 += 2;
215         c0 += 2;
216       }
217       if (nc & 1) {
218         _mm_store_ss(c4, vacc4x0123);
219         _mm_store_ss(c3, vacc3x0123);
220         _mm_store_ss(c2, vacc2x0123);
221         _mm_store_ss(c1, vacc1x0123);
222         _mm_store_ss(c0, vacc0x0123);
223       }
224 
225       nc = 0;
226     }
227   } while (nc != 0);
228 }
229